How a System Finds and Creates Machine Learning Models

This patent describes a system that helps users find existing machine learning models or algorithms for a specific task and, if needed, automatically trains a new model using a selected algorithm.

What it covers

This system acts like a smart librarian for machine learning. First, it registers many machine learning algorithms and their details (metadata) in a special index (Claim 1). When a user needs a machine learning model for an analytics task, they tell the system what they need through a user interface (Claim 1). The system then converts this request into search terms. It first looks for already trained machine learning models that fit the criteria (Claim 1). If it finds none, it then searches its index of *algorithms* to find suitable ones (Claim 1). It shows the user a list of matching algorithms, potentially ranked by how well they fit the request (Claim 3). If the user picks an algorithm, the system then trains a brand new machine learning model using that chosen algorithm, employing standardized tools called universal APIs (Claim 1, Claim 4). For example, if a user needs a model to predict house prices, the system might first check for existing house price prediction models. If none are found, it would then suggest algorithms known for regression tasks, and if the user selects one, it would train a new house price prediction model.

What it doesn't cover

• —Does not cover systems that only store and search for already trained machine learning models without also indexing and suggesting algorithms for new model training.
• —Does not cover systems that only provide a repository for machine learning algorithms without also offering a search engine based on user-specified analytics tasks.
• —Does not cover training machine learning models without the use of a plurality of universal application programming interfaces (APIs) as specified in Claim 1.
• —Does not cover systems that do not first attempt to find existing trained models before searching for algorithms to train new ones, as described in Claim 1.
• —Does not cover systems that do not generate and store metadata models for each registered machine learning algorithm or trained model.

The clever bit

The clever part is the two-tiered search approach: first checking for existing *trained models* and only then, if none are found, searching for *algorithms* to train a new one, all while using universal APIs for consistent training.

Why it matters

This patent addresses a significant challenge in machine learning development: finding or creating the right model for a specific task efficiently. By providing a structured way to search for both existing models and the algorithms to build new ones, it helps streamline the development process. This approach can reduce redundant work and accelerate the deployment of AI solutions across various industries.

Real-world examples

• 1.IBM Watson Studio
• 2.Google Cloud AI Platform
• 3.Amazon SageMaker
• 4.Microsoft Azure Machine Learning
• 5.Enterprise AI/ML platforms